[llvm] [AMDGPU][test] fix the error case in update_mc_test_check script (PR #112731)

[Brox Chen via llvm-commits]

================
@@ -32,7 +31,20 @@ def invoke_tool(exe, cmd_args, testline, verbose=False):
     cmd = 'echo "' + testline + '" | ' + exe + " " + args
     if verbose:
         print("Command: ", cmd)
-    out = subprocess.check_output(cmd, shell=True)
+
+    # if not is used in runline, the command might or might not
+    # return non-zero code on a single line run
+    try:
+        out = subprocess.check_output(cmd, shell=True, stderr=subprocess.DEVNULL)
+    except:
+        if prefix_not:
+            cmd = 'echo "' + testline + '" | ' + "not " + exe + " " + args
+            if verbose:
+                print("Command: ", cmd)
+            out = subprocess.check_output(cmd, shell=True, stderr=subprocess.DEVNULL)
----------------
broxigarchen wrote:

When the runline is start a `not`, the overall file is expected to fail and return a non-zero exit code. However, when the whole test is split and executed line-by-line, each line might, or might not fail and we don't know about it.

Thus I launch two run of the command, one with `not` one without `not`, and  we take the output when it gives zero exit code.

https://github.com/llvm/llvm-project/pull/112731